Grok 4.5

Learning Objectives

• Understand where Grok 4.5 sits in xAI's model lineup and what "Opus-class" actually means
• Read xAI's own benchmark claims critically — where Grok 4.5 leads and where it trails
• Identify when Grok 4.5's token efficiency and low cost are genuine advantages

What Is Grok 4.5?

Grok 4.5 is the flagship model from SpaceXAI — the xAI unit now operating under SpaceX after the February 2026 acquisition — released to the public on July 8, 2026. It is built on the 1.5-trillion-parameter V9 foundation and was trained alongside Cursor, the AI coding editor SpaceX agreed to acquire for $60 billion in June 2026. That pairing shows up in the model's strengths: Grok 4.5 is tuned hard for agentic coding, and it ships as the default model inside Cursor on every plan.

Elon Musk describes Grok 4.5 as an "Opus-class" model — meaning it plays in the same tier as Anthropic's Claude Opus 4.8, "but faster, more token-efficient and lower cost." As a tier description that holds up; as a "beats Opus" claim it is more selective, and the honest reading of xAI's own benchmark chart is below.

Key Capabilities

Built on the V9 Foundation, Trained with Cursor

Grok 4.5 is the first xAI flagship on the V9 foundation (1.5 trillion parameters). Training it alongside Cursor — rather than adapting a general model to coding after the fact — is xAI's bet that a model co-developed with a real coding environment will handle multi-step software tasks more reliably. It runs at roughly 80 tokens per second and is the default model in Grok Build.

Token Efficiency Is the Real Headline

Grok 4.5's most defensible claim is not raw capability but efficiency. On SWE-Bench Pro, xAI reports Grok 4.5 resolves tasks using an average of about 15,954 output tokens, versus roughly 67,020 for Claude Opus 4.8 at maximum effort on the same benchmark — close to a four-times reduction. Fewer tokens per task means lower cost and faster completion for agentic workloads that loop many times, which is exactly where coding agents rack up spend.

Pricing Built for Agentic Loops

At $2 per million input tokens and $6 per million output tokens, Grok 4.5 undercuts most frontier coding models on headline price, and the token-efficiency gap widens the real-world cost advantage further. For teams running agents that make thousands of model calls a day, the combined effect is the model's clearest selling point.

The Honest Benchmark Picture

xAI published a benchmark chart with the launch. Read in full, it shows a split decision against Claude Opus 4.8:

The takeaway: Grok 4.5 is genuinely frontier-competitive on coding, wins clearly on efficiency and cost, and trades wins with Opus on raw task resolution. "Opus-class" is fair; "the best coding model" is not something xAI's own numbers support.

Grok 4.5 vs. Other Frontier Models

Related Tools

• Grok 4.1 — the prior-generation xAI flagship with a 2 million token context window and real-time X data
• Grok — the consumer Grok chat interface, now powered by Grok 4.5
• Claude Opus 4.8 — the Anthropic tier Grok 4.5 benchmarks itself against
• Cursor — the coding editor Grok 4.5 was trained alongside and ships inside

Strengths

• Token efficiency — resolves coding tasks in roughly a quarter of the output tokens Claude Opus 4.8 uses on SWE-Bench Pro, the model's most defensible edge
• Low cost — $2 input and $6 output per million tokens, undercutting most frontier coding models before the efficiency gap is even counted
• Cursor-native — trained alongside Cursor and shipped as the default there on every plan, so it is tuned for real agentic coding rather than benchmarks alone
• Frontier-competitive coding — beats Claude Opus 4.8 on DeepSWE 1.0 and Terminal-Bench 2.1

Limitations and Considerations

• "Opus-class," not "beats Opus" — on xAI's own chart, Claude Opus 4.8 still leads on DeepSWE 1.1 and SWE-Bench Pro task resolution
• Vendor-reported benchmarks — the comparison numbers come from xAI's launch chart, not independent evaluation; treat them as claims until third parties reproduce them
• Closed model — API-only, with no downloadable weights or self-hosting
• Younger enterprise stack — xAI's tooling, support, and third-party ecosystem remain less mature than OpenAI's, Anthropic's, or Google's

Key Takeaways

• Grok 4.5 (July 8, 2026) is SpaceXAI's flagship, built on the 1.5-trillion-parameter V9 foundation and trained alongside Cursor, which SpaceX agreed to acquire for $60 billion
• Elon Musk calls it "Opus-class" — accurate as a tier description; it beats Claude Opus 4.8 on two of four coding benchmarks xAI published and trails on the other two
• Its clearest edge is efficiency: about four-times fewer output tokens than Opus 4.8 on SWE-Bench Pro, at $2 input and $6 output per million tokens
• It is the default model in Grok Build and ships inside Cursor on every plan, running at roughly 80 tokens per second
• The benchmark numbers are xAI's own — competitive and credible, but worth treating as vendor claims until independently reproduced